Two-stage black-box target detection explainability method based on feature gradient

Abstract

The application relates to a two-stage black-box target detection explainability method based on feature gradient, and belongs to the field of trusted deep learning. The method uses the feature gradient idea to disassemble the classification task and the regression task of target detection into two subnetworks for explanation, and obtains the related importance of the classification subnetwork and the regression subnetwork through the feature gradient method. The application can explain the network behavior from the two paths of the classification subnetwork and the regression subnetwork, provide the network basis for each identification result, and thus achieve the purpose of making the black-box model transparent. The method takes the smooth feature gradient as an entrance, finds the concentrated area of network attention, and presents the contribution of the area in a way that is easy for humans to understand, and uses the knowledge in the network to explain the behavior of the network.

Description

Technical Field

[0001] This invention belongs to the field of trusted deep learning and relates to an interpretability method for two-stage black-box object detection based on feature gradients. Background Technology

[0002] Interpretable black-box two-stage object detection models can be applied to fields sensitive to accuracy and with low error tolerance, such as medicine and the military. In medicine, object detection technology is commonly used in fully automated surgical robot systems. The robot uses visual sensors and software running the object detection system to calculate the location of the surgical target and the category of the selected target through model computation. When training the robot's object detection model, using the model's interpretation allows us to check whether the model's learning meets our expectations and whether it truly identifies the target's category based on its intrinsic features rather than inferring from environmental or background features. We can also check whether the features used by the model when finding the target location are reasonable and whether the target location information provided by the features is correct (for example, if the surgical location is the lung, then the features of the navel should provide a vertically upward arrow, because the lung is always directly above the navel). During the training phase, interpretability serves as an indicator of training progress, ensuring the model's reliability. During the inference phase, the object detection model can also generate its interpretation in real time as the surgical robot performs the surgery. This interpretation can be used to monitor the model's performance, checking whether the classification and location-finding decisions made during runtime are correct and whether the correct features are used. Furthermore, the interpretation of classification can provide human doctors with relevant information: the location of the patient's lesion, the type of lesion, and which parts the network inferred the lesion from. Utilizing object detection interpretability during inference not only monitors the model's performance and ensures its reliability but also provides human doctors with additional information for more accurate patient diagnoses.

[0003] Interpretability research on deep black-box networks primarily focuses on the most widely used deep convolutional networks. These interpretability methods mainly address object recognition tasks, aiming to explain the object categories generated by the network from the perspective of its inherent characteristics or mathematical principles. However, in object detection tasks, the network not only needs to infer the category of the object appearing in the input image but also needs to infer the boundary location of the object within the image based on extracted features. Clearly, object detection tasks are far more complex than object recognition tasks. Since object detection networks typically use black-box deep convolutional neural networks as their core, interpretability research on object detection networks is more challenging than that on object recognition networks.

[0004] Based on different interpretation methods, interpretability methods for deep convolutional neural networks can be divided into two main categories: locally interpretable methods and globally interpretable methods. Locally interpretable methods focus on a single input sample, explaining network behavior through features such as importance, rule sets, saliency maps, template features, and example features. Globally interpretable methods, on the other hand, directly study the model parameters, interpreting the network model through methods such as local interpretation sets and model distillation.

[0005] Interpretability research for object detection is relatively scarce. The few existing interpretability methods either require exhaustive enumeration of input features or necessitate training an additional network to extract interpretable features. Therefore, the interpretability efficiency of these methods is relatively low. Furthermore, the explanations of these methods all treat the object detection network as a whole, without separately explaining the classification and regression sub-networks. Users can only know whether the object detection model has problems, but cannot explore the problems more comprehensively or with finer granularity. Some object detection networks employ complex mathematical models to demonstrate network interpretation, setting a barrier to their industrial application and limiting the usability and intuitiveness of interpretability methods.

[0006] Therefore, there is an urgent need to invent a new two-stage black-box object detection interpretability method to solve the following problems: 1) Existing object detection interpretability methods lack independent interpretation of the classification subnetwork and the regression subnetwork; 2) Existing object detection interpretability methods require training additional networks or require a large amount of computation. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide an interpretability method for two-stage black-box object detection based on feature gradients, which respectively demonstrates the reasons for network classification and the reasons for network location of the target, thereby making the black-box object detection model more transparent.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A two-stage black-box object detection interpretability method based on feature gradient is proposed. The method utilizes the concept of feature gradient to decompose the classification and regression tasks of object detection into two sub-networks for interpretation. The correlation importance of the classification sub-network and the regression sub-network are obtained by using the feature gradient method.

[0010] Furthermore, to obtain the relevant importance of the classification subnetwork, that is, to interpret the target recognition classification subnetwork, the specific steps include:

[0011] S101: Train the Faster R-CNN network model using the original training dataset to obtain a Faster R-CNN model that performs well on the dataset.

[0012] S102: Select an original image I to be interpreted from a test dataset that has the same distribution as the training dataset used in step S101 but whose samples do not overlap. Input the image I into the trained Faster R-CNN network model f to obtain the recognition result f(I) = p of the original image sample. I ; where c I It is a classification of targets, p I It is the probability that the target belongs to this category;

[0013] S103: Perturb the input image I using a Gaussian noise model to generate k perturbation samples I1, I2, ..., I i ,…,I k All perturbation samples are input into the trained Faster R-CNN network for inference, resulting in the recognition results f(I1),(I2),…,(I...) for k perturbation samples. i ),…,(I k );

[0014] S104: Treat all (original image samples and perturbation samples) recognition results as dependent variables, and solve for the smooth feature gradient of the original image I; calculate the mean of the gradient matrices of all recognition results to obtain the importance gradient matrix of the classification sub-network.

[0015] S105: Gradient matrix of classification importance Normalization is performed so that the values ​​of the importance matrix are in the range of 0 to 255, making the regions of interest to the network have a darker color;

[0016] S106: Calculate the importance score matrix. Convert it into an image and overlay it onto the original image I to generate the final interpretation of the classification subnetwork.

[0017] Furthermore, in step S103, a Gaussian noise model is used. Perturbation is applied to the original image I, i.e., arbitrary perturbation samples are generated. in, It is the correlation coefficient of noise, which controls the impact of noise on the original image and should prevent the network's recognition results from changing.

[0018] Furthermore, in step S104, the importance gradient matrix of the classifier sub-network is solved using the smooth feature gradient method. The solution formula is as follows:

[0019]

[0020]

[0021] in, c() represents the gradient matrix of the object detection network's output with respect to the input image, I ij Let I represent each pixel in the original image I, and k represent the sample size of the Gaussian noise perturbation.

[0022] Furthermore, in step S105, maximum and minimum value normalization is used to limit the importance values ​​in the matrix to RGB pixel values ​​between 0 and 255. The maximum and minimum value normalization is shown below.

[0023]

[0024] in, It is the maximum value in the matrix, and It is the minimum value in the matrix.

[0025] Furthermore, in step S106, the generated image becomes the network attention point cloud. Regions with darker pixel values ​​in the point cloud are the regions that the classification sub-network focuses on during discrimination, and these appear as darker colors in the original image.

[0026] Furthermore, to obtain the relevance of the regression subnetwork, i.e., to interpret the target location regression subnetwork, the specific steps include:

[0027] S201: Select the same original image I used in step S101, input it into the trained Faster R-CNN network model f, and obtain the recognition result f(I) = [x,y,w,h] of the target location regression subnetwork of the original image sample, where x and y are the horizontal and vertical coordinates of the center point of the target regression box, and w and h are the width and height of the target regression box.

[0028] S202: Scalarize the regression result location information vector to generate the regression location result information scalar g(I);

[0029] S203: Perturb the input image I using a Gaussian noise model to generate k perturbation samples I1, I2, ..., I i ,…,I k All perturbation samples are input into the trained Faster R-CNN network for inference, resulting in the recognition results g(I1), g(I2), ..., g(I) of k perturbation samples. i ),…,g(I k );

[0030] S204: Treat all (original image samples and perturbation samples) recognition results as dependent variables, and solve for the smooth feature gradient of the original image I; calculate the mean of the gradient matrices of all recognition results to obtain the importance gradient matrix of the regression subnetwork;

[0031] S205: Gradient matrix of regression importance Normalization is performed so that the values ​​of the importance matrix are in the range of 0 to 255, making the regions of interest to the network have a darker color;

[0032] S206: Regress the importance score matrix Convert it into an image and overlay it onto the original image I.

[0033] S207: Regression Importance Score Matrix Sort all values ​​in descending order; assume the sorted regression importance score matrix... Sequence M s Let m1, m2, ..., m i ,…,m t ;

[0034] S208: Find sequence M s The largest unprocessed value in the network is used to obtain the RPN (Region Proposal Network) feature region corresponding to the point where the value is located;

[0035] S209: Find the foreground confidence through the classification channel of the RPN feature point, find the basic anchor regression box with the highest confidence, find the final position regression box of the basic anchor regression box through the regression channel, and finally draw regression trend arrows starting from the four corners of the basic anchor point and pointing to the corresponding corners of the final position regression box.

[0036] S210: If the intersection-over-union (IoU) ratio between the base anchor regression box found in step S209 and the already found base anchor regression boxes is greater than η, then ignore this regression box and return to step S207; continue this process until n valid regression boxes are found; where η represents the IoU threshold between a base anchor regression box and the already interpreted base anchor regression region. Base anchor regression boxes exceeding this threshold are considered duplicate interpretations and can be ignored.

[0037] Furthermore, in step S202, the scalarization process can be implemented in two ways: L1-normal method and L2-normal method; the constructed scalarization formula is as follows:

[0038] f(I) = [x, y, w, h]

[0039]

[0040] g(I)=|x|+|y|+|w|+|h|(L1-norm)

[0041] Where 1 represents L1 paradigm, also known as LASSO regression paradigm; L2 represents L2 paradigm, also known as ridge regression paradigm; and norm represents the English word for paradigm.

[0042] Furthermore, in step S208, the coordinates of any point in the original map domain are converted to the coordinates of any point in the RPN feature domain using the following coordinate transformation function:

[0043]

[0044] Where x is the x-coordinate or y-coordinate of the original image domain to be transformed, and s is the scaling factor of the pooling layer in the VGG16 network.

[0045] Furthermore, steps S203, S204, S205, and S206 are the same as steps S103, S104, S105, and S106, and will not be repeated here.

[0046] Furthermore, in step S209, the trend arrows of the basic anchor point regression box explain the hints and contributions of the region selected by the basic regression box to the regression subnetwork for target detection.

[0047] Furthermore, in step S210, the final generated attention point cloud and n regression trend arrow diagrams are the final interpretation of the regression subnetwork.

[0048] The beneficial effects of this invention are as follows: This invention can explain network behavior through two paths—classification subnetwork and regression subnetwork—providing network evidence for each recognition result, thereby achieving the goal of making the black-box model transparent. This method uses smooth feature gradients as an entry point to find the concentrated region of network attention and presents the specific contribution of this region in a way that is easily understood by humans, using the knowledge within the network to explain its behavior. Furthermore, this method does not require a large amount of additional computation or the training of independent network structures throughout the entire explanation process, exhibiting better computational economy compared to existing methods. This invention uses Faster R-CNN as the research model and can be easily transferred to other two-stage object detection models based on Faster R-CNN.

[0049] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0051] Figure 1 This is an overall flowchart of the classification subnetwork interpretation in the method of the present invention;

[0052] Figure 2 This is an overall flowchart of the regression subnetwork interpretation in the method of the present invention;

[0053] Figure 3 This demonstrates the effectiveness of the two-stage black-box target detection network interpretability method based on feature gradients in this invention.

[0054] Figure 4 This study compares the interpretability of target detection network methods with those of mainstream interpretability methods. Detailed Implementation

[0055] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0056] Please see Figures 1-4 This invention provides a two-stage black-box object detection interpretability method based on feature gradients. By utilizing the concept of feature gradients, the classification and regression tasks of object detection are decomposed into two sub-networks for interpretation. The correlation importance of the classification sub-network and the regression sub-network are obtained through the feature gradient method.

[0057] The explanation of the classification subnetwork specifically includes the following steps:

[0058] Step S11: Train the Faster R-CNN network model using the original training dataset to obtain a Faster R-CNN model that performs well on the dataset.

[0059] This embodiment uses the VOC2007 object detection database as an example to explain the specific implementation process in detail. The VOC2007 dataset contains 9963 images, which are divided into training and test sets in an 8:2 ratio. In this embodiment, Faster R-CNN uses VGG16 as the feature extraction network. The Faster R-CNN model is trained using the training set from the VOC2007 dataset, and the model with the highest mAP value on the test set is selected as the experimental model.

[0060] Step S12: Select an original image I to be interpreted from a test dataset that has the same distribution as the training dataset used in step S11 but whose samples do not overlap. Input the image I into the trained Faster R-CNN network model f to obtain the recognition result f(I) = p of the original image sample. I Among them, c I It is a classification of targets, p I It is the probability that the target belongs to this category.

[0061] In this embodiment, a single object detection operation is broken down into two tasks: object classification and object location regression. An input image is arbitrarily selected from the predefined test dataset and fed into the network to obtain the result c of a classification subnetwork. I ,p I The results of a regression subnetwork [x,y,w,h] are given first. The classification subnetwork is explained first.

[0062] Step S13: Perturb the input image I using a Gaussian noise model to generate k perturbation samples I1, I2, ... I i ,…,I k All perturbation samples are input into the trained Faster R-CNN network for inference, yielding the recognition results f(I1), f(I2), ..., f(I...) for k perturbation samples. i ),…,f(I k ).

[0063] In this embodiment, a Gaussian noise model is used. Perturbation is applied to the original image I, i.e., arbitrary perturbation samples are generated. This is the correlation coefficient for noise, which controls the impact of noise on the original image and should prevent changes in the network's recognition results. Gaussian noise is introduced here because the gradient of the network output with respect to the input image has a small probability of fluctuation, leading to inaccurate importance scores. By applying multiple small perturbations to the input image, the gradient fluctuations can be weakened in an average form without changing the network output and decisions, making it as close as possible to the true importance score, ultimately achieving "eliminating fluctuations by increasing perturbation".

[0064] Step S14: Treat all recognition results (original image samples and perturbed samples) as dependent variables, and solve for the smoothed feature gradient of the original image I. Average the gradient matrices of all recognition results to obtain the importance gradient matrix of the classification sub-network. Solve for the importance gradient matrix using the smoothed feature gradient method, as shown below.

[0065]

[0066]

[0067] Step S15: Gradient matrix of classification importance Normalization is performed so that the values ​​of the importance matrix are in the range of 0 to 255, giving the regions of interest to the network a darker color.

[0068] In this embodiment, maximum-minimum normalization is used to limit the importance values ​​in the matrix to the effective RGB pixel values ​​of 0-255, making the matrix easier to identify after conversion into a point cloud image. The maximum-minimum normalization is shown below.

[0069]

[0070] in, It is the maximum value in the matrix, and It is the minimum value in the matrix.

[0071] Step S16: Generate the classification importance score matrix The image is converted into a digital image and overlaid on the original image I to generate the final interpretation of the classification sub-network. The image generated in this step becomes the network's attention point cloud. Regions with darker pixel values ​​in the point cloud are the regions that the classification sub-network focuses on during discrimination, which are represented by darker colors in the original image.

[0072] The interpretation of the regression subnetwork specifically includes the following steps:

[0073] Step S21: Select the same original image I used in step S11 and input it into the trained Faster R-CNN network model f to obtain the recognition result f(I) = [x, y, w, h] of the target location regression subnetwork of the original image sample. Where x and y are the horizontal and vertical coordinates of the center point of the target regression box, and w and h are the width and height of the target regression box.

[0074] Step S22: Scalarize the regression result location information vector to generate the regression location result information scalar g(I).

[0075] In this embodiment, the scalarization process can be implemented in two ways: L1-normal method and L2-normal method. The constructed scalarization formula is as follows:

[0076] f(I) = [x, y, w, h]

[0077]

[0078] g(I)=|x|+|y|+|w|+|h|(L1-norm)

[0079] The L1-normative and L2-normative methods show little difference in the effectiveness of generating interpretations. However, if the direct summation method is used, i.e., g(I) = x + y + w + h, the interpretation effect will be significantly worse. This is because when positive and negative values ​​change in the same direction and at the same rate, they will cancel each other out and cannot be reflected in g(I).

[0080] Step S23: Perturb the input image I using a Gaussian noise model to generate k perturbation samples I1, I2, ... I i ,…,I k All perturbation samples are input into the trained Faster R-CNN network for inference, yielding the recognition results g(I1), g(I2), ..., g(I...) for k perturbation samples. i ),…,g(I k ).

[0081] Step S24: Treat all (original image samples and perturbation samples) recognition results as dependent variables, and solve for the smoothing feature gradient of the original image I. Calculate the mean of the gradient matrices of all recognition results to obtain the importance gradient matrix of the regression subnetwork.

[0082] Step S25: Calculate the regression importance gradient matrix Normalization is performed to ensure the importance matrix values ​​are within the range of 0 to 255, resulting in darker colors for regions of interest to the network. The maximum and minimum value normalization is as follows:

[0083] Step S26: Convert the regression importance score matrix Convert it into an image and overlay it onto the original image I.

[0084] Step S27: For Sort all values ​​in descending order. Assume the sorted values ​​are... Sequence M s Let m1, m2, ..., m i ,…,m t .

[0085] Step S28: Find sequence M s The largest unprocessed value in the network is used to obtain the RPN (Region Proposal Network) feature region corresponding to the point where the value is located.

[0086] In this embodiment, the following coordinate transformation function is used to convert the coordinates of any point in the original map domain to the coordinates of any point in the RPN feature domain:

[0087]

[0088] Where x is the x-coordinate or y-coordinate of the original image domain to be converted, and s is the scaling factor of the pooling layer in the VGG16 network. In this embodiment, s = 16. It should be noted that when VGG16 performs pooling operations, the length and width of the original image are not always integer multiples of the scaling factor s. There are two solutions to this problem: (1) Supplementation method. Fill the pixel regions that are not integer multiples with 0 pixels until they are integer multiples of k. (2) Ignore method. Directly ignore the pixels in the regions that do not meet the requirements. In this embodiment, the backbone network adopts solution (2) for pixels in the regions of the original image that do not meet the requirements of integer multiples. Therefore, the feature region after pooling does not contain the relevant information of the ignored pixels, and it is impossible for them to become important pixels that determine the prediction process of the network. When converting pixels from the original image domain to the RPN feature domain, the problem that the size of the original image is not an integer multiple of k can be ignored.

[0089] Step S29: Find the foreground confidence through the classification channel of the RPN feature point, find the basic anchor regression box with the highest confidence, find the final position regression box of the basic anchor regression box through the regression channel, and finally draw regression trend arrows starting from the four corners of the basic anchor point and pointing to the corresponding corners of the final position regression box.

[0090] In this embodiment, each RPN feature point has corresponding classification and regression information channels. The classification information channel has 18 values, representing the probability of including foreground and background targets within the nine basic anchor regression boxes responsible for that RPN feature point. The regression information channel has 36 values, representing the coordinates and width / height offsets of the regression boxes containing potential targets within the nine basic anchor regression boxes responsible for that RPN feature point. The trend arrows for the basic anchor regression boxes explain the hints and contributions of the area selected by this basic regression box to the regression subnetwork for target detection.

[0091] Step S210: If the intersection-over-union (IoU) ratio of the baseline anchor regression box found in step S29 with the previously found baseline anchor regression boxes is greater than η, then ignore this regression box and return to step S7. Continue this process until n valid regression boxes are found.

[0092] In this embodiment, the basic anchor point regression boxes with similar coverage areas play similar roles. Therefore, we calculated the intersection-union ratio (IUR) of different basic regression boxes and only displayed the regression trends of basic anchor point regression boxes with smaller intersecting areas, i.e., those encompassing as different features as possible. On the attention point cloud, this is represented as clusters of non-intersecting highlighted areas. For the quantization of smaller intersecting areas, consistent with the training of the object detection network, η = 0.7 was set. Furthermore, n = 3 was set because experiments show that the number of highlighted clusters on the attention point cloud generally does not exceed three. It should be noted that a larger n value more completely explains the network behavior. The final generated attention point cloud and three regression trend arrow diagrams represent the final interpretation of the regression sub-network.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A two-stage black-box target detection interpretability method based on feature gradients, characterized in that, Using the concept of feature gradient, the classification and regression tasks of object detection are decomposed into two sub-networks for interpretation, and the correlation importance of the classification sub-network and the regression sub-network are obtained by using the feature gradient method respectively; Obtaining the relevance of the classification subnetwork, i.e., interpreting the object recognition classification subnetwork, specifically includes the following steps: S101: Train the Faster R-CNN network model using the original training dataset to obtain a Faster R-CNN model that performs well on the dataset; S102: Select an original image to be interpreted from a test dataset that has the same distribution as the training dataset used in step S101, but whose samples do not overlap. The input is fed into the trained Faster R-CNN network model. In the process, the recognition results of the original image sample are obtained. ;in, It is the probability that the target belongs to this category; S103: Apply a Gaussian noise model to the input image Perturb and generate k Disturbance samples All perturbation samples are input into the trained Faster R-CNN network for inference, resulting in... k Identification results of individual perturbation samples ; S104: Treat all recognition results as dependent variables, and process the original image... Solve for the smooth feature gradient; average the gradient matrices of all recognition results to obtain the importance gradient matrix of the classification sub-network; S105: Gradient matrix of classification importance Normalization is performed so that the values ​​of the importance matrix are in the range of 0 to 255, making the regions of interest to the network have a darker color; S106: Calculate the importance score matrix. Convert to an image and overlay it onto the original image. The final explanation of the generated classification subnetwork is then provided. Obtaining the relevance of the regression subnetwork, i.e., interpreting the target location regression subnetwork, specifically includes the following steps: S201: Select the original image The input is fed into the trained Faster R-CNN network model. In this process, the target location regression subnetwork of the original image sample is used to obtain the recognition results. ,in, and These are the x and y coordinates of the center point of the target regression box. and These are the width and height of the target regression box; S202: Scalarize the regression result location information vector to generate a regression location result information scalar. ; S203: Use a Gaussian noise model on the input image Perturb and generate k Disturbance samples All perturbation samples are input into the trained Faster R-CNN network for inference, resulting in... k Identification results of individual perturbation samples ; S204: Treat all recognition results as dependent variables, and process the original image... Solve for the smooth feature gradient; average the gradient matrices of all recognition results to obtain the importance gradient matrix of the regression subnetwork; S205: Normalize the regression importance gradient matrix so that the values ​​of the importance matrix are in the range of 0 to 255, so that the regions of interest to the network have a darker color; S206: Convert the regression importance score matrix into an image and overlay it onto the original image. superior; S207: Sort all values ​​in the regression importance score matrix in descending order; assume the sorted regression importance score matrix sequence for ; S208: Find the sequence The largest unprocessed value in the range is used to obtain the RPN feature region corresponding to the point where the value is located; S209: Find the foreground confidence through the classification channel of the RPN feature point, find the basic anchor regression box with the highest confidence, find the final position regression box of the basic anchor regression box through the regression channel, and finally draw regression trend arrows starting from the four corners of the basic anchor point and pointing to the corresponding corners of the final position regression box. S210: If the intersection-union ratio (IUU) of the baseline anchor point regression box found in step S209 with the already found baseline anchor point regression boxes is greater than... Then ignore this regression box and return to step S207; continue this process until the regression box is found. There are 10 valid regression boxes; among them... This represents the cross-union threshold between a given baseline regression bounding box and the interpreted baseline regression region.

2. The two-stage black-box target detection interpretability method according to claim 1, characterized in that, In step S103, a Gaussian noise model is used. In the original image Perturbation is performed on the sample, i.e., arbitrary perturbation samples are generated. ;in, It is the correlation coefficient of noise.

3. The two-stage black-box target detection interpretability method according to claim 2, characterized in that, In step S104, the importance gradient matrix of the classifier sub-network is solved using the smooth feature gradient method. The solution formula is as follows: in, This represents the gradient matrix of the object detection network's output with respect to the input image. This represents each pixel in the original image I. k This represents the sample size for Gaussian noise perturbation.

4. The two-stage black-box target detection interpretability method according to claim 1, characterized in that, In step S202, the scalarization process is implemented in two ways: L1-normal method and L2-normal method; the constructed scalarization formula is as follows: in, This represents the L1 paradigm, also known as the LASSO regression paradigm. This refers to the L2 paradigm, also known as the ridge regression paradigm. norm The English word for paradigm.

5. The two-stage black-box target detection interpretability method according to claim 1, characterized in that, In step S208, the following coordinate transformation function is used to convert the coordinates of any point in the original map domain to the coordinates of any point in the RPN feature domain: in, The x-coordinate or y-coordinate of the original map domain to be converted. This is the scaling factor for the pooling layer in the VGG16 network.

Images